Pavement disease real-time detection and maintenance evaluation method and system based on visual intelligence

Abstract

The invention relates to the technical field of pavement disease real-time detection and maintenance evaluation methods, in particular to a pavement disease real-time detection and maintenance evaluation method and system based on visual intelligence, and the method comprises the steps: obtaining to-be-detected image data; obtaining normalized data; constructing a multi-dimensional feature matrix set based on the normalized data; performing dynamic weighted combination on the feature matrix set to generate a combined feature matrix; executing nonlinear collaborative enhancement convolution based on the combined feature matrix to obtain an optimized feature output matrix so as to improve the precision of disease detection; based on the optimized feature output matrix, applying adaptive dynamic loss optimization, and generating disease detection information; according to the method, the to-be-maintained road surface is detected, the maintenance evaluation result of the to-be-maintained road surface is generated based on the disease detection information, and layer-by-layer progressive and collaborative enhancement of the features are realized in the disease detection process through a multi-level feature convolution and nonlinear collaborative enhancement convolution technology, so that high detection precision can still be kept under small sample data and complex disease features.

Description

Technical Field

[0001] The present invention relates to the technical field of real-time detection and maintenance assessment methods for pavement defects, and more specifically, to a real-time detection and maintenance assessment method and system for pavement defects based on visual intelligence. Background Art

[0002] In the development of today's road infrastructure, the detection and maintenance of pavement defects are particularly important. With the increase in traffic volume and climate change, road defects such as cracks, potholes, and subsidence are becoming increasingly serious, posing severe challenges to traffic safety and pavement quality. In order to effectively solve these problems, existing technologies have gradually developed a series of pavement defect detection methods based on image recognition. However, these traditional methods still have major technical defects, especially in application scenarios with complex defect types, fuzzy features, and high real-time detection requirements.

[0003] Most of the existing pavement defect detection technologies rely on single-layer or multi-layer convolution models based on convolutional neural networks (CNNs), such as the YOLO series and FasterR-CNN. These models usually perform well on simple defect images, but their performance drops significantly when faced with complex, noisy actual pavement scenes. The main reason is that the existing methods have limitations in the hierarchy and refinement of feature extraction, and are not adaptable enough to defect features. For example, the edge features of defects such as cracks and potholes are often subtle and fuzzy, which places higher requirements on the accuracy of feature extraction. However, due to the lack of multi-scale aggregation and adaptive optimization of features, existing methods are often unable to accurately identify these defect details, thereby reducing the accuracy of detection.

[0004] At the same time, the existing technology usually adopts fixed convolution kernels and hierarchical designs in the convolution processing process, and lacks a mechanism for dynamically adjusting the importance of different features. Single-layer convolution or fixed feature extraction methods can easily lead to low sensitivity of the model to a few features, especially in disease detection of small samples and unbalanced data, which can easily miss key features. This unbalanced performance of features limits the robustness of the detection system for different disease types. In addition, most traditional methods fail to introduce a dynamic weight adjustment mechanism in the loss function, making it impossible for the detection model to be adaptively optimized according to the actual situation of different disease types and classification errors. Therefore, in practical applications, these methods perform poorly on rare or blurred-edge disease types, and the detection accuracy is not ideal.

[0005] Another major problem is the contradiction between real-time and evaluation accuracy of existing methods. Due to complex convolution calculations and feature extraction operations, traditional models often have difficulty meeting real-time requirements when detecting large-scale road defects. Inefficient detection speed not only affects the practicality of the system, but may also lead to missed detection of defects due to delays. In addition, traditional methods also lack a dynamic adjustment mechanism when generating maintenance assessment results. The assessment results usually rely on fixed thresholds and are difficult to adapt to changes in the actual road environment. As a result, it is difficult for existing technologies to provide timely and accurate references for disease maintenance in actual road detection, affecting the reliability of maintenance assessments. Summary of the invention

[0006] In response to the above technical problems, the present invention proposes a real-time pavement disease detection and maintenance assessment method based on visual intelligence, aiming to overcome the shortcomings of the existing technology in complex disease detection, unbalanced feature extraction, and the contradiction between real-time and assessment accuracy.

[0007] The present invention provides a real-time detection and maintenance assessment method for pavement damage based on visual intelligence, comprising: Acquire image data to be detected; Performing normalization processing on the image data to obtain normalized data; Constructing a multidimensional feature matrix set based on the normalized data to represent the disease features in the image; Performing dynamic weighted combination on the feature matrix set to generate a combined feature matrix; Performing nonlinear collaborative enhancement convolution based on the combined feature matrix to obtain an optimized feature output matrix to improve the accuracy of disease detection; Applying adaptive dynamic loss optimization based on the optimized feature output matrix to generate disease detection information; and A maintenance assessment result of the road surface to be maintained is generated based on the disease detection information.

[0008] Preferably, the step of obtaining the image data to be detected includes: collecting image data in the maintenance area through the vehicle terminal, normalizing it to a uniform resolution and scale, and forming an image matrix ,in, is the image matrix to be detected, is the number of rows in the image, is the number of columns of the image, and is based on the image matrix Forming a normalized data matrix , the normalized data matrix is ​​used for subsequent feature extraction and analysis.

[0009] Preferably, based on the normalized data matrix Obtain a feature matrix set, the feature matrix set including a feature vector set ,in, For the eigenvectors, corresponding to the eigenmatrix ; According to the significance coefficient of each eigenvector , forming a feature weight set , where the weight calculation formula is: ; in, is a set of feature vectors, containing multi-dimensional features; For the feature vectors; is the feature matrix; is the feature weight set; is the feature significance coefficient, and the feature matrix set is used as input for feature combination and optimization.

[0010] Preferably, based on the feature matrix set and the weight set Get the combined feature matrix , the combined feature matrix The formula for generating is: ; in, It is a combined feature matrix, which contains a comprehensive representation of multiple features; is the noise balance parameter; is a noise matrix; the combined feature matrix is ​​used for the nonlinear collaborative enhancement convolution step.

[0011] Preferably, the nonlinear collaborative enhancement convolution comprises the following steps: Based on the combined feature matrix Using convolution kernel matrix Get the convolution result matrix , the formula is: ; in, is the combined feature matrix, is the convolution kernel matrix, is the convolution result matrix, is a convolution operator; based on the convolution result matrix , use the adaptive weight balancing formula to obtain the weight adjustment factor , the formula is: ; in, is the weight adjustment factor, is the adaptive coefficient, is the feature vector, is the feature mean; based on the weight adjustment factor And the convolution result matrix , the inverse feature enhancement function is used to obtain the enhancement matrix , the formula is: ; in, is the enhancement matrix, is the reverse enhancement coefficient, is a small constant, Represents the elements of the feature matrix after nonlinear collaborative enhancement convolution, specifically the eigenvalue of a certain position obtained after convolution of the input image feature matrix; Use nonlinear suppressed convolution on the enhancement matrix Perform inhibition and generate inhibition matrix , the formula is: ; in, is the inhibition matrix, is the activation function, is the suppression coefficient; finally based on the suppression matrix and weight adjustment factor Get the optimized feature output matrix , the formula is: ; in, is the optimized feature output matrix, used in the dynamic loss optimization step, is the sum of feature dimensions.

[0012] Preferably, based on the optimized feature output matrix Get adaptive dynamic loss value , the adaptive dynamic loss value is used to evaluate the disease detection effect of the model, and the loss value formula is: ; in, is the loss value, is the actual disease distribution, To predict disease distribution, is the dynamic weight factor, is a small constant, Represents the total number of samples in the disease detection process, that is, the number of disease instances used to calculate the loss value.

[0013] Preferably, based on the adaptive dynamic loss optimization result, the disease detection information in the image to be detected is obtained, and the disease detection information includes the disease location, disease type and disease degree.

[0014] A real-time road disease detection and maintenance assessment system based on visual intelligence includes: an on-board terminal for collecting image data, and a cloud server for data processing and analysis, wherein the on-board terminal is used to collect images to be detected in a maintenance area and upload the image data to the cloud server; the cloud server generates a feature matrix set based on the image data, and dynamically weights and combines the feature matrix set to obtain a combined feature matrix; the cloud server performs multi-level collaborative enhancement convolution and dynamic loss optimization based on the combined feature matrix to obtain disease detection information, and generates a maintenance assessment result of the road surface to be maintained based on the disease detection information.

[0015] Preferably, the cloud server includes: an image preprocessing unit, a multidimensional feature extraction unit, a feature combination unit, a nonlinear collaborative convolution unit and a dynamic loss optimization unit. The nonlinear collaborative convolution unit is used to perform multi-level collaborative feature enhancement convolution based on an optimized feature combination matrix, and the dynamic loss optimization unit is used to obtain optimized disease detection information based on an adaptive dynamic loss function.

[0016] Preferably, the cloud server also includes a data storage unit, a data acquisition unit and a calculation unit, the data acquisition unit is used to collect the defect location, the data storage unit is used to store the coordinate data and type information of the defect location, and the calculation unit is used to calculate the defect spacing according to the defect location, and generate the maintenance assessment result of the road surface to be maintained based on the defect spacing.

[0017] The present invention has the following beneficial effects: First, the present invention can accurately extract disease features at different scales and achieve synergy and superposition between features by constructing a multi-dimensional feature matrix set and dynamically weighted combination, effectively solving the detection blind spot problem caused by feature imbalance in traditional methods. Through multi-level feature convolution and nonlinear collaborative enhancement convolution technology, the present invention achieves layer-by-layer progression and collaborative enhancement of features in the disease detection process, thereby maintaining high detection accuracy under small sample data and complex disease features.

[0018] In addition, the present invention uses adaptive dynamic loss optimization to give the model the ability to adjust the sensitivity of different types of diseases, so that rare or blurred disease information is not ignored during detection. This adaptive dynamic loss optimization can not only effectively adjust the detection weights of various diseases, but also adjust the loss function in real time according to the classification error, making the detection process more accurate. While realizing dynamic optimization of the detection process, the present invention further ensures the improvement of detection speed, can achieve efficient real-time detection in complex road environments, and generate accurate maintenance assessment results based on disease detection information.

[0019] Through the above-mentioned technical improvements, the present invention achieves a balance between detection accuracy and real-time performance, and forms effective synergy in feature extraction and model optimization. Specifically, the multidimensional feature extraction method and adaptive dynamic loss optimization of the present invention work together to successfully solve the performance degradation problem of traditional models when detecting diverse diseases. The optimized model can not only capture multi-level disease detail features, but also generate maintenance assessment results that meet actual needs under the condition of efficient operation, thereby improving the reliability and practicality of the entire system.

[0020] In summary, the present invention demonstrates significant technical advantages in real-time detection and maintenance assessment of pavement defects through an innovative combination of feature extraction, convolution enhancement, and loss optimization, making up for the shortcomings of existing technologies in detection accuracy, speed, and adaptability, and providing a set of efficient and accurate intelligent solutions for road maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow logic block diagram of the method of the present invention.

[0022] Figure 2 It is a logical block diagram of the system composition of the present invention.

[0023] Figure 3 It is a logic block diagram of the image preprocessing and feature extraction of the present invention.

[0024] Figure 4 It is a logic block diagram of the feature combination and convolution processing of the present invention.

[0025] Figure 5 A logical block diagram of the dynamic loss optimization and evaluation of the present invention. DETAILED DESCRIPTION

[0026] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0027] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0028] Please refer to Figure 1-5The invention proposes a real-time detection and maintenance assessment method for pavement defects based on visual intelligence, which includes the following steps: obtaining image data to be detected, normalizing the image data to obtain normalized data of a uniform scale, and then constructing a multidimensional feature matrix set based on the normalized data to represent the pavement defect characteristics in the image. Through a dynamic weighted combination method, the feature matrix set is used to generate a combined feature matrix. Then, nonlinear collaborative enhancement convolution is performed based on the combined feature matrix to obtain an optimized feature output matrix, and the optimized feature output matrix is ​​used to improve the accuracy of defect detection. Finally, adaptive dynamic loss optimization is applied based on the optimized feature output matrix to generate defect detection information, and a maintenance assessment result of the pavement to be maintained is generated based on the defect detection information.

[0029] Preferably, in the step of normalizing the image data, the image data can be unified to a preset scale, such as 640x480 pixels, by unifying images of different resolutions and sizes. This standard size helps to improve the speed and consistency of image processing in subsequent steps. In experience, controlling the image resolution within this range can ensure clarity without placing an excessive load on system performance. The purpose of normalizing the data is to ensure the consistency of the input data format and to provide a stable data basis for the construction of a multidimensional feature matrix.

[0030] According to a preferred embodiment of the present invention, the step of acquiring the image data to be detected includes using the vehicle-mounted terminal 1 to collect image data in the maintenance area and normalizing the data to a uniform resolution and scale. The obtained image data forms an image matrix ,in, is the number of rows in the image, is the number of columns of the image. Based on this, a normalized data matrix is ​​further generated. , in order to facilitate subsequent feature extraction and analysis.

[0031] Preferably, the vehicle-mounted terminal 1 can be equipped with a high-performance camera to reduce image blur caused by vehicle motion through a built-in stabilization algorithm to ensure data reliability. Such a configuration not only improves the accuracy of image acquisition, but also better preserves detail features during the conversion process of the image matrix. In order to reduce errors, the present invention can set a pixel deviation with a threshold of 0.05 to detect image blur, thereby automatically reshooting or reprocessing the image to ensure data accuracy.

[0032] According to a preferred embodiment of the present invention, based on the normalized data matrix Obtain a feature matrix set. The feature matrix set includes a feature vector set ,in, For the eigenvectors, corresponding to the eigenmatrix The significance coefficient of each eigenvector is Used to calculate the feature weight set , and preferably according to the formula: ; in, The statistical significance of each eigenvector is used to form a set of weights. Represents the image matrix under different features. Specifically, The choice of can be based on empirical values ​​and float in the range of [0.1,1]; when the significance of a feature in the data set exceeds the threshold of 0.6, its weight can be increased so that the feature occupies a larger proportion in subsequent analysis. Such a setting can effectively enhance the expression ability of disease features in different dimensions, and ultimately enhance the adaptability and accuracy of the detection model.

[0033] According to a preferred embodiment of the present invention, based on the feature matrix set and the feature weight set , will generate a combined feature matrix The formula for generating the combined feature matrix is: ; in, is a combined feature matrix, which represents the integration of multiple features. is a noise balance parameter, preferably between 0.05 and 0.1, to ensure that the noise effect is suppressed while maintaining effective features. is the noise matrix. Combined feature matrix Used as input for nonlinear collaborative enhancement convolution. Preferably, when the noise distribution of the combined feature matrix is ​​higher than 0.08, it will be automatically rebalanced to ensure that the noise matrix does not interfere with the expression of the main disease features. In practical scenarios, the combined feature matrix can take into account multiple feature dimensions to improve the aggregate expression of pavement disease information.

[0034] According to a preferred embodiment of the present invention, the execution process of the nonlinear collaborative enhancement convolution is as follows: based on the combined feature matrix , first use the aggregate convolution kernel Get the convolution result matrix , the specific formula is: ; in, is the combined feature matrix, is the convolution kernel matrix, is the convolution result matrix, is a convolution operator. This operation can aggregate similar damage information in different feature matrices, and is particularly suitable for pavement damage detection with complex feature details. In the convolution process, in order to avoid excessive convolution leading to loss of details, it is preferred to set the convolution kernel size to 3x3.

[0035] Then, the weight adjustment factor is calculated by the adaptive weight balancing formula : ; in, is the weight adjustment factor, is the adaptive coefficient, preferably in the range of [0.5, 1.5], which is used to balance the influence of the feature. is the feature vector, By setting the threshold value to 0.6, we can ensure that the weight is automatically adjusted when the feature significance exceeds the threshold value, so as to improve the significant expression of the disease feature.

[0036] Further, based on the weight adjustment factor And the convolution result matrix , the inverse feature enhancement function is used to obtain the enhancement matrix : ; in, is the enhancement matrix, is the reverse enhancement coefficient. The empirical value is usually between 0.1 and 0.2 to ensure that the strength of reverse enhancement is moderate. is a small constant, preferably 0.001, to avoid the denominator being zero. Through reverse enhancement, the system can achieve a balance between high-frequency and low-frequency features, effectively improving feature contrast; Represents the elements of the feature matrix after nonlinear collaborative enhancement convolution, specifically the eigenvalue at a certain position obtained after convolution of the input image feature matrix.

[0037] Finally, a nonlinear suppressed convolution is used on the enhancement matrix Perform inhibition and generate inhibition matrix : ; in, is the inhibition matrix, is the activation function, preferably the ReLU function to improve the nonlinear effect, is the suppression coefficient, preferably between 0.5 and 1, to ensure that noise is controlled during feature amplification. and weight adjustment factor Get the optimized feature output matrix , the formula is: ; is the sum of feature dimensions. This optimized feature output matrix can better represent pavement damage information and provide effective support for the generation of subsequent damage assessment results.

[0038] According to a preferred embodiment of the present invention, the optimized feature output matrix obtained is Finally, the present invention further optimizes the accuracy of disease detection by calculating the adaptive dynamic loss function. The formula for the loss value used to evaluate the disease detection effect of the model is as follows: ; in, is the loss value, Indicates the actual distribution of the disease. is the disease distribution predicted by the model, is a dynamic weight factor. Preferably, the threshold is set to 0.05 to avoid calculation distortion caused by excessive or too small weight. is a small constant used to prevent the denominator from being zero. The empirical value is 0.001. Represents the total number of samples in the disease detection process, that is, the number of disease instances used to calculate the loss value.

[0039] Adaptive dynamic loss function is particularly important in disease detection because it can adjust the model's performance on different disease types. For example, when detecting rare disease types, by increasing the dynamic weight factor The value of increases the detection accuracy of the model on rare diseases, while the weight of common disease types is reduced, thus achieving a balance between different disease types. This method not only reduces the overall error, but also enhances the stability of the model, enabling it to maintain efficient and accurate detection results in different disease scenarios.

[0040] According to a preferred embodiment of the present invention, based on the adaptive dynamic loss optimization result, the disease detection information in the image to be detected is obtained. The disease detection information includes the location, type and severity of the disease. Preferably, in order to improve the positioning accuracy of the disease information, the GPS module is used to additionally obtain the geographic coordinates of the location of the disease, and the disease information is associated with the geographic coordinates. The severity of the disease can be judged by combining the detected area and type of the disease. For example, the severity of the crack can be evaluated by its length or width, while the potholes are graded by their depth. In addition, to ensure the accuracy of the classification, the disease detection confidence threshold can be set to 0.8, and the disease information below this value will not be included in the final result to avoid misjudgment.

[0041] In order to present the disease information more intuitively, the detection information can be fed back to the user interface in the form of image tags, and the location information, type and severity of the disease can be displayed on the electronic map to form an intuitive disease distribution map. This not only facilitates the rapid understanding of road conditions, but also distinguishes the type and severity of the disease by color grading, thereby improving the work efficiency of road maintenance personnel.

[0042] According to a preferred embodiment of the present invention, the present invention also discloses a hardware system structure, namely, a real-time detection and maintenance assessment system for pavement defects based on visual intelligence. The system includes a vehicle-mounted terminal 1 and a cloud server 2, wherein the vehicle-mounted terminal 1 is used to obtain image data to be detected and upload the image data to the cloud server 2, and the cloud server 2 is used to analyze the image data, generate disease detection information, and generate a maintenance assessment result of the road surface to be maintained based on the detection information. Preferably, the vehicle-mounted terminal 1 may include hardware modules such as a camera, a GPS module, and a data processing unit to facilitate image acquisition and location marking. At the same time, the cloud server 2 is configured with a high-performance processor and sufficient storage space to process the uploaded image data in real time.

[0043] During the image acquisition process, the vehicle terminal 1 can take pictures of the road ahead regularly or intermittently to ensure full coverage of the road section to be inspected. The shooting frequency can be automatically adjusted according to the driving speed. For example, when the vehicle speed is less than 20 kilometers per hour, an image is taken every 5 seconds; when the vehicle speed is higher than 50 kilometers per hour, the shooting frequency can be increased to one every 3 seconds to reduce missed inspections. Through the image data uploaded in real time by the vehicle terminal 1, the cloud server 2 can process and analyze the image data, and automatically feed back the disease detection results to the maintenance department so that repair measures can be taken in time.

[0044] According to a preferred embodiment of the present invention, the cloud server 2 includes multiple functional units, specifically including an image preprocessing unit, a multidimensional feature extraction unit, a feature combination unit, a nonlinear collaborative convolution unit, and a dynamic loss optimization unit. The image preprocessing unit normalizes the image data uploaded by the vehicle terminal 1 to ensure the consistency of image data from different sources; the multidimensional feature extraction unit constructs a feature matrix set based on the image data and generates the weight of each feature; the feature combination unit generates a combined feature matrix based on the feature matrix set and the weight, and further inputs the nonlinear collaborative convolution unit for collaborative feature enhancement. The nonlinear collaborative convolution unit mainly realizes multi-level feature combination through convolution and weight balance adjustment, thereby improving the adaptability and accuracy of the model in disease detection. The dynamic loss optimization unit optimizes the detection results based on the adaptive dynamic loss function, so that the disease detection information is more accurate, especially when detecting complex disease structures.

[0045] In specific applications, the image preprocessing unit usually sets the preprocessing threshold to 0.05 pixel deviation during normalization operation to reduce the impact of image resolution on feature extraction. At the same time, in order to further improve the effect of the feature combination unit, the present invention can preferably use the ReLU activation function in the nonlinear collaborative convolution unit to increase the nonlinear expression ability of the model. The collaborative work of these feature units significantly improves the accuracy of disease detection in complex road environments.

[0046] According to a preferred embodiment of the present invention, the cloud server 2 further includes a data storage unit, a data acquisition unit and a computing unit for recording the disease detection information and calculating the maintenance assessment results. Specifically, the data acquisition unit collects the location data of the disease and transmits it to the data storage unit, and the storage unit saves the coordinate data and type information of the disease location to ensure that each piece of disease information detected can be recorded in detail for subsequent analysis and reference. The computing unit further calculates the disease spacing based on the stored data to generate the maintenance assessment results of the road surface to be maintained.

[0047] In order to improve the accuracy of the assessment, the present invention can set a threshold of 5 meters when calculating the distance between defects, so as to judge the clustering of defects and give priority to marking the areas formed by adjacent defect locations. In addition, the calculation unit can also formulate different maintenance levels based on the type and severity of the defect. For example, when the severity of the defect in the same area exceeds the preset threshold of 80%, the system automatically marks the area as "urgent repair required"; if the concentration of the defect is less than 40%, it is marked as "routine inspection". Such maintenance assessment results can more accurately reflect the actual condition of the road surface to be maintained, thereby providing a scientific and reasonable decision-making basis for road maintenance work.

[0048] Through the collaborative work of the above-mentioned functional units, the real-time detection and maintenance assessment system of pavement defects based on visual intelligence of the present invention realizes intelligent management of the whole process from image data acquisition, feature extraction, defect analysis to the generation of maintenance assessment results. In practical applications, the present invention can provide reliable detection information in different pavement defect scenarios, and effectively improve the accuracy and efficiency of defect detection through adaptive optimization and collaborative calculation. The following are the embodiments and comparative examples of the present invention, which specifically test the innovative points of the present invention, that is, the superiority of the real-time detection and maintenance assessment method of pavement defects based on visual intelligence in actual detection. Through the application of data sets and detailed comparison of indicators, the significant improvement of the present invention in detection accuracy, detection speed and maintenance assessment accuracy is further verified. The embodiments of this article are tested based on professional pavement defect data sets (such as Crack500 data sets and self-built multi-class disease data sets), and the test results and their significance are described in detail from multiple perspectives.

[0049] In an embodiment, the method of the present invention is applied to a test data set, which contains multiple categories of disease images, including typical road surface disease types such as cracks, potholes, and subsidence. The method of the present invention adopts the comprehensive processing flow of feature extraction, collaborative convolution enhancement, and adaptive dynamic loss function optimization, and completes the processing in a cloud server to ensure the efficiency of data processing and the accuracy of the model. Specific detection indicators include three key parameters: detection accuracy, detection speed, and maintenance assessment accuracy.

[0050] In order to evaluate the actual effect of the method, the test of the embodiment uses the following indicators: 1. Detection accuracy: This indicator indicates the accuracy of disease detection and is calculated as the ratio of the number of correctly detected disease images to the total number of disease images. This indicator directly reflects the detection performance of the model.

[0051] 2. Detection speed: Measured in frames per second (FPS), it indicates the detection speed of the system under actual road conditions. A higher detection speed indicates that the model can process image data in real time and efficiently.

[0052] 3. Maintenance assessment accuracy: The degree of match between the maintenance assessment results generated based on the test results and the manually labeled results is expressed as an accuracy percentage. The higher the value, the closer the system assessment results are to the actual situation.

[0053] In this embodiment, the optimization scheme of the present invention is selected, and the recognition and evaluation effect of the model on different disease types is improved through multi-dimensional feature extraction, collaborative convolution and adaptive dynamic loss function optimization. The threshold is set in the experiment to ensure the reliability of the accuracy of disease detection and maintenance assessment accuracy.

[0054] The comparative example uses a traditional YOLOv5 model without feature combination and convolution enhancement, and mainly extracts and classifies disease features through a single-layer convolution without dynamic loss optimization. This method is still based on the same image data set for detection, but due to the lack of complex feature combination and adaptive optimization strategy, the detection performance is reduced in terms of accuracy and evaluation. By comparing the detection results of different methods, it can be clearly seen that the performance improvement of the present invention.

[0055] Table 1 shows the main test results in the examples and comparative examples, including the test accuracy, test speed and maintenance assessment accuracy. The test results show that the test method of the present invention is superior to the comparative model in all key indicators.

[0056] Table 1 Comparison of pavement disease detection results project Example Comparative Example Detection accuracy (%) 92.5 78.3 Detection speed (FPS) 25 20 Maintenance assessment accuracy (%) 91.0 76.8 From the test results in Table 1, it can be seen that the present invention has significant advantages in the accuracy of driving safety index: Detection accuracy: In terms of detection accuracy, the accuracy of the embodiment reached 92.5%, which is significantly higher than 78.3% of the control example. The present invention uses multidimensional feature extraction, weight balance and collaborative convolution enhancement in the detection process, so that the disease features can be effectively extracted and expressed at multiple scales, thereby improving the detection accuracy of the disease. Especially for small cracks and diseases with blurred edges, the characteristics of collaborative convolution enable the model to identify more accurately, while the traditional YOLOv5 model performs poorly when dealing with complex disease edges.

[0057] Detection speed: In terms of detection speed, the detection speed of the embodiment reaches 25FPS, which is higher than 20FPS of the comparative example. This result is due to the optimization of the multi-dimensional feature combination of the present invention, which reduces the repeated calculation of feature extraction and further accelerates the image processing speed through collaborative convolution. For real-time road detection applications, the higher frame rate ensures that road surface disease information can be fed back in time during driving, allowing maintenance personnel to respond quickly.

[0058] Maintenance assessment accuracy: In terms of maintenance assessment accuracy, the embodiment achieved 91.0%, much higher than the 76.8% of the comparative example. The method of the present invention can more accurately assess the type and severity of pavement damage, especially after the use of adaptive dynamic loss optimization, the system can perform hierarchical processing for different types of damage, significantly improving the accuracy of maintenance assessment. In contrast, the comparative model is prone to ignoring rare types of damage due to the lack of dynamic loss optimization, resulting in deviations in the assessment results.

[0059] In summary, the present invention effectively improves the accuracy of pavement disease detection and the reliability of maintenance assessment through a series of optimization technologies. These test results fully demonstrate the superiority of the present invention in terms of detection accuracy, real-time performance and assessment accuracy, providing more scientific and reliable technical support for road maintenance.

[0060] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A real-time pavement disease detection and maintenance assessment method based on visual intelligence, characterized in that: include: Acquire image data to be detected; Performing normalization processing on the image data to obtain normalized data; Constructing a multidimensional feature matrix set based on the normalized data to represent the disease features in the image; Performing dynamic weighted combination on the feature matrix set to generate a combined feature matrix; Performing nonlinear collaborative enhancement convolution based on the combined feature matrix to obtain an optimized feature output matrix to improve the accuracy of disease detection; Applying adaptive dynamic loss optimization based on the optimized feature output matrix to generate disease detection information; as well as A maintenance assessment result of the road surface to be maintained is generated based on the disease detection information.

2. The method for real-time detection and maintenance assessment of pavement defects based on visual intelligence according to claim 1 is characterized in that: The steps of obtaining the image data to be detected include: collecting image data in the maintenance area through the vehicle terminal, normalizing it to a uniform resolution and scale, and forming an image matrix ,in, is the image matrix to be detected, is the number of rows in the image, is the number of columns of the image, and is based on the image matrix Forming a normalized data matrix , the normalized data matrix is ​​used for subsequent feature extraction and analysis.

3. The method for real-time detection and maintenance assessment of pavement defects based on visual intelligence according to claim 2 is characterized in that: Based on the normalized data matrix Obtain a feature matrix set, the feature matrix set including a feature vector set ,in, For the eigenvectors, corresponding to the eigenmatrix ; According to the significance coefficient of each eigenvector , forming a feature weight set , where the weight calculation formula is: ； in, is a set of feature vectors, containing multi-dimensional features; For the feature vectors; is the feature matrix; is the feature weight set; is the feature significance coefficient, and the feature matrix set is used as input for feature combination and optimization.

4. The method for real-time detection and maintenance assessment of pavement damage based on visual intelligence according to claim 3 is characterized in that: Based on the feature matrix set and the weight set Get the combined feature matrix , the combined feature matrix The generation formula is: ； in, It is a combined feature matrix, which contains a comprehensive representation of multiple features; is the noise balance parameter; is a noise matrix; the combined feature matrix is ​​used for the nonlinear collaborative enhancement convolution step.

5. The method for real-time detection and maintenance assessment of pavement defects based on visual intelligence according to claim 4 is characterized in that: The nonlinear collaborative enhancement convolution comprises the following steps: Based on the combined feature matrix Using convolution kernel matrix Get the convolution result matrix , the formula is: ； in, is the combined feature matrix, is the convolution kernel matrix, is the convolution result matrix, is a convolution operator; based on the convolution result matrix , use the adaptive weight balancing formula to obtain the weight adjustment factor , the formula is: ； in, is the weight adjustment factor, is the adaptive coefficient, is the feature vector, is the feature mean; based on the weight adjustment factor And the convolution result matrix , the inverse feature enhancement function is used to obtain the enhancement matrix , the formula is: ； in, is the enhancement matrix, is the reverse enhancement coefficient, is a small constant, Represents the elements of the feature matrix after nonlinear collaborative enhancement convolution, specifically the eigenvalue of a certain position obtained after convolution of the input image feature matrix; Use nonlinear suppressed convolution on the enhancement matrix Perform inhibition and generate inhibition matrix The formula is: ； in, is the inhibition matrix, is the activation function, is the suppression coefficient; finally based on the suppression matrix and weight adjustment factor Get the optimized feature output matrix , the formula is: ； in, is the optimized feature output matrix, used in the dynamic loss optimization step, is the sum of feature dimensions.

6. The method for real-time detection and maintenance assessment of pavement damage based on visual intelligence according to claim 5 is characterized in that: Based on the optimized feature output matrix Get adaptive dynamic loss value , the adaptive dynamic loss value is used to evaluate the disease detection effect of the model, and the loss value formula is: ； in, is the loss value, is the actual disease distribution, To predict disease distribution, is the dynamic weight factor, is a small constant, Represents the total number of samples in the disease detection process, that is, the number of disease instances used to calculate the loss value.

7. The method for real-time detection and maintenance assessment of pavement defects based on visual intelligence according to claim 1 is characterized in that: Based on the adaptive dynamic loss optimization result, the disease detection information in the image to be detected is obtained, and the disease detection information includes the disease location, disease type and disease degree.

8. A real-time road damage detection and maintenance assessment system based on visual intelligence, characterized by: include: A vehicle-mounted terminal for collecting image data, and a cloud server for data processing and analysis, wherein the vehicle-mounted terminal is used to collect images to be detected in the maintenance area and upload the image data to the cloud server; the cloud server generates a feature matrix set based on the image data, and dynamically weights and combines the feature matrix set to obtain a combined feature matrix; The cloud server performs multi-level collaborative enhancement convolution and dynamic loss optimization based on the combined feature matrix to obtain disease detection information, and generates a maintenance assessment result of the road surface to be maintained based on the disease detection information.

9. The real-time pavement disease detection and maintenance assessment system based on visual intelligence according to claim 8 is characterized in that: The cloud server includes: an image preprocessing unit, a multidimensional feature extraction unit, a feature combination unit, a nonlinear collaborative convolution unit and a dynamic loss optimization unit. The nonlinear collaborative convolution unit is used to perform multi-level collaborative feature enhancement convolution based on an optimized feature combination matrix, and the dynamic loss optimization unit is used to obtain optimized disease detection information based on an adaptive dynamic loss function.

10. The real-time pavement disease detection and maintenance assessment system based on visual intelligence according to claim 8 is characterized in that: The cloud server also includes a data storage unit, a data acquisition unit and a calculation unit. The data acquisition unit is used to collect the location of the defect, the data storage unit is used to store the coordinate data and type information of the location of the defect, and the calculation unit is used to calculate the distance between the defects according to the location of the defect, and generate the maintenance assessment result of the road surface to be maintained based on the distance between the defects.

Images